Stem splitter: split any song into 4 stems, free

SplitVocals uses Demucs v4, a hybrid transformer model trained specifically to separate music into four sources: vocals, drums, bass, and everything else (guitars, keys, pads, synths — grouped as "other"). The model analyzes the full mix and reconstructs each source as its own audio stream, rather than just filtering frequencies, so instruments that overlap in pitch still come out as distinct stems.

The whole pipeline — decoding your file, running the model, and re-encoding the output — happens on your own CPU or GPU through WebAssembly or WebGPU. There's no server in the loop, so there's no upload step and no queue.

Producers and DJs pull the drum or bass stem from a reference track to build a mashup, or lift an acapella to lay over a new beat. Remix artists use isolated stems as raw material without re-recording anything. Sample-based musicians dig individual instrument lines out of old records for new productions. Because output stems are yours to use in original work you have the rights to, they drop straight into a DAW project.

Feed it the highest-quality source you have — a WAV or FLAC gives the model more detail than a heavily compressed MP3. Dense mixes with lots of reverb or many instruments sharing the same frequency range are the hardest case for any separation model, so expect some bleed between stems (a little vocal breath in the "other" stem, for example) on those tracks. Simpler, well-mixed arrangements split the most cleanly.

Each stem downloads as WAV or 320k MP3, and you can grab all four at once as a ZIP.

Cloud-based splitters typically meter you by the minute and make you wait for a render queue to clear. Running the model on your own hardware removes both limits: no per-minute cost, and processing starts the moment you drop the file. On a GPU-accelerated browser like Chrome or Edge, a full song typically finishes in one to three minutes; Safari and Firefox use a CPU fallback that takes longer but produces the same output.